Multiparametric MRI Tumor Probability Model for the Detection of Locally Recurrent Prostate Cancer After Radiation Therapy: Pathologic Validation and Comparison With Manual Tumor Delineations

Abstract

Purpose: Focal salvage treatments of recurrent prostate cancer (PCa) after radiation therapy require accurate delineation of the target volume. Magnetic resonance imaging (MRI) is used for this purpose; however, radiation therapy-induced changes complicate image interpretation, and guidelines are lacking on the assessment and delineation of recurrent PCa. A tumor probability (TP) model was trained and independently tested using multiparametric magnetic resonance imaging (mp-MRI) of patients with radio-recurrent PCa. The resulting probability maps were used to derive target regions for radiation therapy treatment planning.

Methods and materials: Two cohorts of patients with radio-recurrent PCa were used in this study. All patients underwent mp-MRI (T2 weighted, diffusion-weighted imaging, and dynamic contrast enhanced). A logistic regression model was trained using imaging features from 21 patients with biopsy-proven recurrence who qualified for salvage treatment. The test cohort consisted of 17 patients treated with salvage prostatectomy. The model was tested against histopathology-derived tumor delineations. The voxel-wise TP maps were clustered using k-means to generate a gross tumor volume (GTV) contour for voxel-level comparisons with manual tumor delineations performed by 2 radiologists and with histopathology-validated contours. Later, k-means was used with 3 clusters to define a clinical target volume (CTV), high-risk CTV, and GTV, with increasing tumor risk.

Results: In the test cohort, the model obtained a median (range) area under the curve of 0.77 (0.41-0.99) for the whole prostate. The GTV delineation resulted in a median sensitivity of 0.31 (0-0.87) and specificity of 0.97 (0.84-1.0) with no significant differences between model and manual delineations. The 3-level clustering GTV and high-risk CTV delineations had median sensitivities of 0.17 (0-0.59) and 0.49 (0-0.97) and specificities of 0.98 (0.84-1.00) and 0.94 (0.84-0.99), respectively.

Conclusions: The TP model had a good performance in predicting voxel-wise presence of recurrent tumor. Model-derived tumor risk levels achieved sensitivity and specificity similar to manual delineations in localizing recurrent tumor. Voxel-wise TP derived from mp-MRI can in this way be incorporated for target definition in focal salvage of radio-recurrent PCa.